Glossary

Memorization

Memorization means that a model has learned a training text so that it later reproduces it word for word or almost word for word. It is a special case of learning, not a separate store: the model holds no text file, but its parameters can be set so that a continuation fits one particular sentence exactly.

Memorization is not always harmful. Common phrases or formulas should be reproducible word for word. It becomes a problem with texts that may not be freely reused, and with names, phone numbers or addresses.

An example: A poem appears on several web pages. A model that has seen it many times can continue the next lines word for word when it is given the first ones.

Not to be confused with a copy: Memorization is not the same as a stored file. Whether it legally counts as a reproduction is an open question that courts will decide. For the assessment, it helps to separate collecting the texts, training and the model’s answer.

Where you’ll come across it: In research on training data, in court cases about AI training, and in reports on whether a model can disclose personal details.

Introduced in “Training Data: Copyright and Privacy” in the section on the model. How the parameters get their settings is explained in “What an AI Model Actually Is”.

Last changed on · commit 653da06